Perception-Action Based Object Detection From Local Descriptor Combination And Reinforcement Learning

This work proposes to learn visual encodings of attention patterns\nthat enables sequential attention for object detection in real world\nenvironments. The system embeds a saccadic decision procedure in\na cascaded process where visual evidence is probed at informative\nimage locations. It is based on the extraction of information theoretic\nsaliency by determining informative local image descriptors that\nprovide selected foci of interest. The local information in terms\nof code book vector responses and the geometric information in the\nshift of attention contribute to recognition states of a Markov decision\nprocess. A Q-learner performs then performs search on useful actions\ntowards salient locations, developing a strategy of action sequences\ndirected in state space towards the optimization of information maximization.\nThe method is evaluated in outdoor object recognition and demonstrates\nefficient performance